Fix the environment
Record host, power mode, runtime version, model artifact and context setting.
Research libraryPerformance claims should be
SelbsAI test register
Performance claims should be
reproducible or absent.
Every published run records the host, processor, memory, accelerator, runtime, model revision, quantisation, checksum, context, prompt, test date and repetitions. Estimates and vendor figures do not enter this register.
Version
1.0
Protocol date
2026-08-09
Protocol published before results.
Protocol01 02 03 04
One protocol, disclosed before the numbers.
The register separates prompt processing from answer generation and keeps the complete environment beside the result.
Warm once
Load the model and complete one unreported warm-up request.
Run five times
Use the versioned prompt and record native runtime counters for every repetition.
Publish the median
Report median prompt and generation rates with all raw run identifiers retained.
Required record
A speed number without these fields is not comparable.
Host · CPU · accelerator
Memory · power mode
Model · revision · checksum
Runtime · quantisation
Context · prompt · output
Rates · latency · repetitions
Verified runsRead the hardware sizing guide →
Published SelbsAI test runs.
No verified physical-host runs are published yet.
This is deliberate. The register is live, the schema is ready and the test runner now emits the required fields. Results will appear only after a named host completes the protocol and the record passes review.
Publication gate
Four rules keep the table honest.
- No estimated or affiliate-supplied results.
- Exact model revision, quantisation and checksum required.
- At least five measured repetitions after one warm-up run.
- Failures and thermal or memory constraints remain in the notes.